Skip to content

Use unfused SDPA for short sequences (q_len <= 128 or kv_len <= 128) #18651

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
kimishpatel wants to merge 9 commits into
base: gh/kimishpatel/224/base
Choose a base branch
from
+5 −1
Open

Use unfused SDPA for short sequences (q_len <= 128 or kv_len <= 128) #18651

Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
d70fe8e
Use unfused SDPA for short sequences (q_len <= 128 or kv_len <= 128)
kimishpatel Apr 1, 2026
1083d69
Update on "Use unfused SDPA for short sequences (q_len <= 128 or kv_l…
kimishpatel Apr 6, 2026
b08de1b
Update on "Use unfused SDPA for short sequences (q_len <= 128 or kv_l…
kimishpatel Apr 6, 2026
1a34cd4
Update on "Use unfused SDPA for short sequences (q_len <= 128 or kv_l…
kimishpatel Apr 7, 2026
11b9a55
Update on "Use unfused SDPA for short sequences (q_len <= 128 or kv_l…
kimishpatel Apr 8, 2026
4b9da2b
Update on "Use unfused SDPA for short sequences (q_len <= 128 or kv_l…
kimishpatel Apr 9, 2026
6acda1b
Update on "Use unfused SDPA for short sequences (q_len <= 128 or kv_l…
kimishpatel Apr 13, 2026
98ebd08
Update on "Use unfused SDPA for short sequences (q_len <= 128 or kv_l…
kimishpatel Apr 13, 2026
e4ba4cf
Update on "Use unfused SDPA for short sequences (q_len <= 128 or kv_l…
kimishpatel Apr 14, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 5 additions & 1 deletion extension/llm/custom_ops/op_sdpa.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,3 @@
/*
* Copyright (c) Meta Platforms, Inc. and affiliates.
* All rights reserved.
Expand Down Expand Up @@ -412,7 +412,11 @@
InvalidArgument,
output);

bool use_unfused_sdpa = seq_len == 1;
// Quantized GEMM kernels may not handle non-contiguous per-head strides
// correctly when seq_dim=ONE and seq_len > 1, so keep the conservative
// condition for quantized inputs.
bool is_quantized = q.scalar_type() == ScalarType::Char;
bool use_unfused_sdpa = (!is_quantized) && (seq_len <= 128 || num_keys_for_causal_attention <= 128);
if (use_unfused_sdpa) {
ET_SWITCH_FLOAT_TYPES(output.scalar_type(), ctx, "sdpa", CTYPE, [&] {
sdpa::impl::cpu_sdpa<CTYPE>(
Expand Down
Loading